Experimental Analysis and Characterization of a WirelessSensor Network Environment

Bogdan PavkovicGrenoble Informatics Laboratory (LIG)

University of Grenoblepavkovic@imag.fr

Fabrice TheoleyreCNRS, LSIIT

University of Strasbourgtheoleyre@unistra.fr

Dominique BarthelOrange Labs, Meylan

dominique.barthel@orange-ftgroup.com

Andrzej DudaGrenoble Informatics Laboratory (LIG)

University of Grenobleduda@imag.fr

ABSTRACTExisting testbeds, even though rare and specialized, are

often not used to their full potential. Collected data duringan experimentation is usually used to evaluate some specifictested aspect. To further benefit from the knowledge gath-ered on a testbed and to obtain the insight into the WSNenvironment itself, we propose a thorough statistical analy-sis. Some of our analysis include radio link characterization,its correlation with environmental parameters as well as aninsight into network dynamics from the point of view of anode and a link. We also discuss how testbeds should bedesigned or improved to provide more detailed informationnecessary for an advanced analysis.

Categories and Subject DescriptorsC.2.1 [Computer-Communication Networks]: [Net-

work Architecture and Design - Wireless Communication]

General TermsExperimentation, Algorithms, Performance

Keywordswireless sensor networks, testbeds, characterization, sta-

tistical analysis

1. INTRODUCTIONWireless Sensor Networks have attracted a lot of atten-

tion for a few years. Most of the research e!ort has fo-cused on the two main issues: routing [1] and power-energysavings [2]. However, the research community has becomeaware that models of wireless multihop networks are too

Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.PE-WASUN’10, October 17–18, 2010, Bodrum, Turkey.Copyright 2010 ACM 978-1-4503-0276-0/10/10 ...$10.00.

simplistic and lead to misleading conclusions. In particular,di!erent simulators have been proven to provide di!erentresults [3]. The radio model especially has a strong impacton performance [4].

To advance the evaluation of various protocols, we can set-up an ad hoc testbed to compare the simulation results withmeasurements gathered on the testbed. However, the mea-surements usually concern only a limited number of testedaspects and setting up operational testbeds requires largehuman e!orts.

Developing new testbeds to accelerate prototyping andto make their evaluation easier has become an interestingresearch objective. Orbitlab [5] proposed for example to de-ploy a testbed in a dynamically reconfigurable grid. Otherauthors provide guidelines to design feasible protocols: theyadvocate for the principle of the simplest is the best [6].However, these commendable e!orts usually do not providegeneric results to the networking research community. Forinstance, they do not consider many important aspects suchas: What are the characteristics of the WSN radio topol-ogy? What is the reliability of a WSN? Are the propertiesstable or do they exhibit some variability or periodicity?We propose here to address one part of these fundamentalconcerns.

In the past, statistical analysis has been applied to tra"canalysis [7] or anomaly detection [8] to extract some cor-relations and salient features. We propose in this paper touse this mathematical approach to characterize the perfor-mances and properties of a WSN. Our analysis includes inparticular:

• characterization of radio links in a WSN: their relia-bility and the correlation between their properties;

• analysis of the network dynamics: how does a WSNchange in time?

• how can we predict the quality of a radio link with alocal and simple measure?

• how can we discard measurement errors?

2. TESTBED DESCRIPTIONWe used a testbed originally designed for validating a

routing protocol [9]. It was composed of 36 Coronis nodes

Figure 1: Deployed topology in an urban environ-ment

implementing the Wavenis technology [10]: a fast frequencyhopping is implemented to be robust to narrow band inter-ference. Nodes operate in the 868 MHz license-free band,emitting at 25 mW with maximum transmission rate of19200 bps. The MAC layer follows a CSMA-CA approachfor medium access contention. Besides, two nodes actedas sinks with a direct connection to the Internet and adatabase for storing received packets. Nodes were deployedover the area of the technical park of Orange Labs in Meylan,France, both indoor and outdoor. Their location is diver-sified enough (e.g. walls, barrier, trees, ceiling) so that awide range of situations is observed. We analyzed the mea-surements of 18 days of operation. Figure 1 presents thedeployed topology in the urban environment.

The testbed was mainly used to validate a routing protocolbased on virtual coordinates: each node maintains a metricrelated to its virtual distance to the sink [11]. The next hopis chosen as the neighbor that is virtually the closest to thesink.

Nodes perform a neighborhood discovery every 13 minutesand maintain a proactive neighborhood table including thevirtual distance and RSSI of each neighbor.

Each node generates a new data packet every 17 minutes.This packet is transmitted in anycast: any sink can be usedto reach the wired part of the network. In order to selectthe next hop (the node that has the lowest virtual distance),a node only has to walk its neighborhood table.

The routed packets, aside from the control fields (sourceand destination ID, sequence number, etc.), contain debuginformation consisting of complete neighborhood tables (theneighbor ID and the received RSSI value: 32 possible lev-els between -108 dBm and -60 dBm in 1.5 dBm increments)and the application payload consisting of measurements ofthe temperature, humidity, and light sensors at the instantjust before sending the packet. Packets successfully receivedat sink nodes were labeled with a timestamp and stored ina database. Table 1 sums-up the important testbed infor-mation.

3. METHODOLOGY

Environment type UrbanNode position Indoor & outdoorSensor type Coronis Wavenis

Number of nodes (sinks) 36(2)Duration of the experiment 18 days

Neigh. discovery period 13 min.Data packet generation period 17 min.

Table 1: Testbed parameters

Link occurrence ratio [% of experiment total length]

Fre

quency

0 5 15 25 35 45 55 65 75 85 950

10

20

30

50

100

Figure 2: Distribution of link occurrence ratio

3.1 Database descriptionTo allow meaningful interpretation and easy use of dif-

ferent types of measured values contained in the receivedrouting packets, the database is divided into few tables:

• the node ID and its geographical position (known priorto deployment) to obtain the geographical topology.We can compare it to the radio topology;

• neighborhood information (neighbor ID and a RSSIvalue). We can observe in particular duration andquality of the links;

• sensor measurements (e.g. humidity, temperature).

On the average, each node sent 1,500 data packets (max-imum sample size) to sinks, where just the ones successfullyarrived were saved in the database. To perform an accuratestatistical analysis, we need to discard received data sam-ples with insu"cient cardinality. Thus, we have removed allthe data samples that count less than 1% of the maximumsize (i.e. 15 entries). They correspond to isolated or faultynodes.

3.2 Bidirectional and Unidirectional linksWe can distinguish between unidirectional and bidirec-

tional links (RSSI measures are available for one or for bothdirections). We obtained 16 unidirectional links and 280bidirectional links.

We define as link occurrence ratio the number of appear-ances of a candidate node in the neighborhood table of a

reference node divided by the total number of tables for thatreference node. In other words, it represents the percentageof the cases where a link between two nodes was detectedand qualified with an RSSI value. We can notice in Figure 2that a significant number of links (20%) exist less than 1% ofthe time. By filtering these sets with too small a cardinality,we in particular eliminated all the unidirectional links: theirdata sets accounted only for 1 to 4 occurrences. Thus, oneof our first results is that the testbed did not have any uni-directional links. However, some of the bidirectional linkscan be asymmetrical (i.e. their quality is di!erent for bothdirections), as explained in one of the following section.

Unidirectional links may appear when antennas are notperfectly omnidirectional [12], the filters are not well-con-ceived [13] or when the nodes do not use the same trans-mission power [14]. Consequently, we can conclude that theCoronis nodes are robust and the hardware is well conceivedand industrialized (i.e. di!erent nodes have the same char-acteristics).

3.3 Filtering dataSince we focus on experimental data, we have to discard

ambiguous measures (i.e. possible outliers or impracticablevalues) to obtain unbiased results. We propose to detect anddiscard this kind of values.

Formally, we consider that a value is an outlier, if it con-forms to the following condition:

x < Q1! 1.5 · IQR " x > Q3 + 1.5 · IQR (1)

where Q1 represents first quantile of observed data set, Q3third quantile and IQR di!erence between them i.e. inter-quantile range.

Thus, we discard all the values that are single isolatedoutliers: only one value is extreme, corresponding surely toa transient behavior. On the contrary, multiple consecutiveoutliers could arise from temporary obstacles (e.g. a deliverytruck, a car) for radio propagation, climatological changes(heavy rain that disturbs radio transmissions and increasesthe humidity measures). Thus, we keep all multiple consec-utive outliers. In other words, we consider that the extremevalues that last for more than 17 minutes are valid. We willgive more attention to multiple consecutive outliers later inthe article to infer the main causes and consequences.

After filtering our experimental dataset, we proceed withthe analysis.

4. LINK QUALITYThe progress in the radio chip design positively impacted

the performance and reliability of WSNs [15]. This mo-tivated us to further investigate the possibility to use theRSSI value as a reliable link quality indicator.

4.1 Radio link SymmetryWe measured the RSSI value in both directions for each

radio link (Figure 3). In this graph, we did not remove thelinks with a very small number of values (as explained insection 3.3) because we aim here at analyzing the reason oftheir existence.

When the points are close to the diagonal, the links aresymmetrical: the quality is identical in both directions. Thereader can remark that contrary to the literature, symmetryis predominant.

!60 !70 !80 !90 !100 !110

!60

!70

!80

!90

!100

!110

Link from N!>M RSSI value[dBm]

Lin

k f

rom

M!

>N

RS

SI

va

lue

[dB

m]

!

!

!

!

!!

!

!

!

!

! !

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

! !!!

!

!

!

!

!

!!

!

!

!

!

! !

!

!

!

!!

!

!

!

!

!!

!

Figure 3: Symmetry of existing bidirectional links

This means that nodes use the same transmission power.Besides, they are also homogenous: the radio hardware be-haves identically. For instance, the radio chips of two di!er-ent radio modules follow the same frequency selectivity (i.e.filters are identical).

Radio links are seldom asymmetrical: these outliers ap-pear for links with a duration less than 1% of total lengthof the experiment. For these rare cases, the quality in onedirection is significantly di!erent, i.e. greater than 10%,sometimes even 55%. This unbalanced representation justi-fies the removing of links with too small cardinalities.

4.2 RSSI distributionTo predict the link behavior with a local and simple met-

ric, the measure should follow a known probability distribu-tion model: we would be able to accurately infer the averagequality of the link by analyzing in real-time the measuredvalues.

The Normal (or Gaussian) distribution is extensively usedsince it models well many natural phenomena, especially forradio propagation (e.g. the Additive White Gaussian Noise).We aim here at verifying if the RSSI measured for each ofthe existing links follows this distribution.

We applied the Shapiro-Wilk test [16], to the measuredRSSI samples. For 92% of the links, the p-value of theShapiro-Wilk test was significantly less than 0.05 and forthe rest barely over this value. This signifies that we needto reject the null-hypothesis, meaning that the RSSI doesnot follow a normal distribution. This corroborates someindoor results [17] and even in outdoor conditions for LOSradio links, the RSSI does not follow a normal distribution.

We compared also these RSSI samples to other two well-known distributions: Logistic and Cauchy. These distribu-tions are the single ones that permit to have this kind ofvalues (close to a log-normal law but with minor variations).

We used a Kolmogorov-Smirnov test, permitting to com-pare a well-known distribution to a collection of samples.

RSSI value [dBm]

De

nsity

0.0

00

.04

0.0

80

.12

!108 !96 !90 !84 !78 !72

Figure 4: Distribution of RSSI value for one of therepresentative link

More precisely, a collection of values are generated accord-ing to the well-known distribution, with the same cardinalityas the set we want to compare to. Then, the Kolmogorov-Smirnov test lets us know if the two collection of valuesfollow the same distribution. For both distribution casesand for all of the links resulting p-value was always close to0, meaning that we have to drop null-hypothesis i.e. RSSIsamples do not follow neither Logistic neither Cauchy dis-tribution. Nevertheless, RSSI distribution that we tried todescribe, has bell shape with high central peak but it isslightly skewed to one side (Figure 4). Thus, no well-knowndistribution permits to have a generic model for such RSSIvalues.

We now aim at demonstrating that the RSSI of the di!er-ent links follow the same distribution. Since they do not fol-low the Normal distribution we have chosen one of the mostfamiliar non-parametric test—the Wilcoxon-Mann-WhitneyTest [16]. Now the null-hypothesis is that values from thetwo independent samples come from the same distribution.The average of p-value for this test was 0.0416 while 89% ofvalues were smaller than 0.05. The null-hypothesis is validsince this p-value is lower than alpha level 0.05. In otherwords, we conclude that the corresponding pairs of samplesdo follow the same distribution.

This can be also observed in Figure 5. We plotted the BoxPlots for some representative links. Even though medianvalues are not perfectly aligned, we can notice that interquantile ranges are similar as well as the skewness of dataand max/min values.

4.3 RSSI periodicityWe also analyzed the di!erence in radio link quality dur-

ing working hours (8am-7pm) and night periods (9pm-6am).Plotting the values we obtained almost the same graph asthe one plotted in Figure 3 showing that links did not changetheir properties during di!erent periods of day. Thus, move-ments of people and vehicles in the technical park duringworking hours do not have any significant impact on theRSSI. RSSI is stable and transmissions are quite robust tosome environment properties changes.

In other words, the PHY layer in the Wavenis nodes is

!1

10

!1

00

!9

0!

80

!7

0!

60

RS

SI valu

e[d

Bm

]

Figure 5: Box plot showing that bidirectional linksfollow the same distribution (with similar parame-ter’s values

robust to interferences because it uses frequency hopping.Moreover, the PHY channel is stable.

4.4 RSSI vs. Link occurrence ratioTo have a more detailed insight into the link occurrence ra-

tio property shown in Figure 2, we tried to observe whetherit can be correlated with the RSSI value.

Figure 6 shows Box Plots for all recognized links in thetestbed separated in 4 groups according to the range of theirlink occurrence ratio without sorting them in ascending or-der by the same criteria.

Looking at this figure we can notice that there is no ev-ident correlation between RSSI value and link occurrenceratio since Box Plot of RSSI covers whole extent of possiblevalues in di!erent link occurrence ratio ranges. However, wecan remark the following points:

• a single RSSI distribution for a particular link doesnot permit to conclude on the occurrence ratio for thislink. Individual conclusions are not possible;

• if we take a closer look at the graph, we can remarkthat each category exhibits di!erent RSSI spreadings.In other words, we could derive a probability of linkoccurrence ratio for di!erent RSSI values. However,this constitutes a global (and not individual) behavior,i.e. RSSI is not directly a good quality estimator;

• for the first range of link occurrence ratio (1-40%),mean value of the RSSI for all of the links do not passabove -90 dBm. Thus, a poor link means obligatory alow RSSI;

• the largest RSSI values mean in most cases we facestable links.

4.5 RSSI vs. sensor measurementsFirst, we checked the correlation between the measured

humidity and the RSSI. During the experiments, nodes hap-pened to be exposed to humidity levels between 0 and 100%relative humidity (RH). We computed the Pearson’s corre-lation factors [16] for all bidirectional links. In all cases, thevalue did not exceed 0.5, plus we have neither a negativenor a positive correlation between the two variables. In the

!110

!100

!90

!80

!70

!60

RS

SI

va

lue

[d

Bm

]

Link occurrence ratio < 40% Link occurrence ratio (40,60) % Link occurrence ratio (60,80) % Link occurrence ratio (80,100) %

Figure 6: Impact of the RSSI value on link occurrence ratio

same way, the correlation is not significant if we only focuson outdoor radio links.

The second test was an attempt at further exploring thecorrelation between humidity and RSSI values, but just tak-ing into account the impact of the extreme (maximum) val-ues of humidity measurements. We extracted subset of top25% of all humidity values (more than 75% of RH) mea-sured during the experiment. Afterward, we computed thedi!erence between the mean value of RSSI for the observedlink and the current value of RSSI at the same temporaryinstant as the measured extreme value of humidity.

If the humidity does not impact the RSSI, we would havean average di!erence equals to 0. This means that the RSSIvalue has the same chance to be greater or lower than theaverage RSSI value. Since we observed this behavior (valuesare very closed to 0, with a varying sign), we can for sure saythat humidity has no impact on the RSSI measurements.

In conclusion, the fast frequency hopping technique is ef-ficient to avoid interferences and frequency-selective fading.

5. NETWORK DYNAMICSThe radio channel is intrinsically unstable, since it is easily

influenced by various environmental parameters. This cre-ates a certain dynamism in the network, where links can eas-ily disappear or re-appear. To optimize the performances,the deployed MAC, topology control and routing protocolsshould self-adapt to changes. We will now focus on the net-work dynamics to understand how it could further impactthe higher layers.

5.1 Neighborhood variationWe first studied the variation in the neighborhood table.

The same remarks hold for all the nodes, and we focus hereon one node, selected randomly. We plotted in Figure 7 thevariation of its neighborhood size.

In the current testbed, there is no hello packets becauseit implements a all-reactive solution. When a node wantsto transmit a data packet, it sends a RTS. All its neigh-bors reply with a CTS including the received RSSI. Thus, anode is able to reconstruct the list of its neighbors and thecorresponding RSSI.

It varies most of the time with rare stable periods that lastat most few samples. We have recognized this behavior as ageneral trend for all the nodes. This raises the question ofwhether a proactive approach is the most accurate solutionfor discovering its neighbors. Indeed, proactive maintenancemay result in ine"cient routing decisions when choosing non

reliable nodes: they can be chosen as next hop because anhello was previously received although they will not cor-rectly receive the next data packet. Although RSSI maybe stable, the radio link may not be. This result tends toconclude that opportunistic solutions, where the next hopis chosen only when the data packet is transmitted, is morerelevant in this environment. Since the next hop is chosenreactively among the nodes that received the data packet,unreliability problem is reduced.

5.2 Link EvolutionAlthough the neighborhood table changes continuously, a

group of stable neighbors may practically exist. In partic-ular, can a large neighbor’s stability be correlated with e.g.the RSSI or the distance between the transmitter and thereceiver?

We have noticed that stable radio links have one of thefollowing properties:

• a high RSSI value (superior to -75 dBm);

• a pair of nodes within one fifth of the radio range(#70 m) and having a medium value of RSSI (between-75 dBm and -90 dBm).

By combining distance and RSSI information, we shouldbe able to predict the link stability. Moreover, there wasno single case in which neighbors with a high value of RSSIwere not among most stable neighbors.

High RSSI could be used as reliable indicator of link stabil-ity when using Wavenis chips [10]. Geographical informationis a plus to cope with medium RSSI values. Similar obser-vation was made by other authors [15] for di!erent type ofradio chips. Nevertheless, there is still substantial free spacein order to make tighter conclusions about the link behaviorwith RSSI in a gray zone (low levels close to the threshold)since it is influenced by various e!ects (multipath, fading,interference, etc.) for which the impact varies over time andaccording to the situation.

5.3 Multiple consecutive outlier distributionAs previously stated in section 3.3 we have kept multiple

consecutive outliers since they depict transitory e!ect thatinfluence the quality of radio channel for a short period. Anoutlier will introduce a bias in the average and median valuesif we do not discard them.

We aim here at analyzing this phenomenon in a moreglobal way, i.e. for all observed links in the network. Thus,we extracted the Cumulative distribution function (CDF) of

!02

46

810

Time

Ne

igh

bo

rho

od

siz

e

Jun 25 Jun 30 Jul 05 Jul 10

!!!!

!

!

!!

!

!

!!

!!

!!!!!!

!!

!!

!!

!

!!

!!!

!

!!!!

!

!

!

!!!

!!!!!

!!!!!!

!!

!

!!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!!

!!

!

!!!

!!

!!

!

!!!!

!

!!

!!

!

!

!

!

!!

!

!

!

!!!

!

!!!

!

!

!

!!!

!

!!

!

!

!

!

!

!!

!

!

!!

!!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!!

!!!

!

!

!

!

!!

!!!

!

!!

!

!

!

!

!

!

!

!

!

!

!

!

!!

!

!!!!

!

!!

!!

!!!!!

!

!!!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!!!

!!

!!

!

!

!!

!

!!

!!!

!

!!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!!

!

!!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!

!!!

!

!

!!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!

!

!

!!

!

!!

!!

!

!!

!

!!

!

!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!!

!

!

!

!

!!

!

!

!!

!!!!

!!

!

!

!

!!

!!

!!

!

!

!!

!

!

!!

!

!

!

!!

!!

!

!

!

!

!

!!!

!

!

!

!

!

!

!

!!!!

!!!

!

!

!

!

!

!

!

!!!

!

!

!

!

!

!

!

!!

!!

!!

!

!

!!

!

!

!

!

!

!

!

!

!

!

!!

!

!

!!

!

!

!

!!

!!

!!

!!!

!

!

!

!

!!

!

!

!!

!

!!

!

!

!

!!

!!!!!!

!

!

!

!

!!

!!!!!!!

!!

!

!

!

!!

!

!!

!

!

!!

!!!

!!!!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!!!!!

!

!

!

!

!!!

!

!

!!!

!

!!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!!

!!

!

!

!

!!

!

!

!!!

!

!!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!!

!!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!!!

!

!

!!

!!

!

!

!!

!

!

!

!

!

!

!

!

!

!

!

!

!!!

!

!

!

!!

!

!

!

!

!!

!

!

!

!

!

!!!

!

!

!!

!!!!

!

!

!

!!

!

!

!!

!!

!

!

!

!

!

!

!

!!!!!

!

!

!!

!

!

!

!

!!

!

!!!

!

!!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!!!

!

!

!

!!!

!

!

!!

!

!

!

!

!

!

!

!!

!

!

!!

!

!!

!!

!

!

!

!

!

!

!

!

!

!

!

!!!!

!

!!

!

!

!

!

!!

!!

!

!

!!!

!

!

!

!

!!

!!

!

!!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!!!

!

!

!

!

!

!

!!

!!

!

!!

!

!

!

!!!

!!

!

!

!

!

!!

!!!

!

!

!!

!!

!!!

!

!!

!!

!

!!

!

!!

!!

!

!

!!!!

!

!

!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!

!!!

!

!

!!!

!

!

!

!

!

!

!

!!

!

!

!

!

!!

!

!

!!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!!

!

!

!!!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!!

!

!

!

!

!!!!

!

!

!

!!!

!

!

!

!!

!!

!

!

!!

!

!

!

!

!!

!

!

!

!

!

!

!

!

!

!

!!!

!

!

!!!

!

!

!

!

!

!

!!

!!

!

!

!

!

!

!

!

!

!

!!

!!

!

!

!

!

!!

!

!

!!

!!

!!

!!

!

!

!

!

!

!!

!

!

!

!!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!!

!

!

!

!!

!!

!

!!!!

!

!

!

!

!

!

!

!

!

!!

!

!

!!

!

!

!!

!

!

!

!!

!

!

!!!

!

!!

!!!!!!!!

!

!

!

!

!

!!

!

!

!!

!

!!!

!

!

!!

!

!!

!

!!

!

!!!

!

!

!

!

!

!

!

!

!

!

!

!

Figure 7: Variation in the neighborhood size for one of the nodes

Number of multiple consecutive outliers

Pro

ba

bili

ty

!

!

!

!

!

!

!

!

!

!! ! ! !!!!!!!!!

!!!!!!!!!!! !!! ! !! !!! !

!

1 2 3 4 8 12 20 30 42 78 138

0.0

0.2

0.4

0.6

0.8

1.0

Figure 8: CDF of multiple consecutive outlier

the number of multiple consecutive outliers. We plotted theresults in Figure 8.

In our static testbed we have 4 or less consecutive outliersin 75% of the cases (the blue circle in Figure 8). In thesame way, for 90% of the cases, we have 8 or less multipleconsecutive outliers (dashed line in the figure). This meansthat we can consider multiple consecutive outliers lastingup to 8 periods as transitory e!ects which interrupt stableradio link for short period. Since observing more than 8consecutive outliers is very rare in a static testbed, we canconsider that this phenomenon is related to a permanenttopology change in a mobile/changing testbed (e.g. buildingmodification).

Let’s still focus on the 4 consecutive outliers case. Wehave approximately 12% of the samples that last for exactly4 outliers (75% - 63%)), and only 5% of th samples thatlast for exactly 5 outliers (80% - 75%). In other words,when a node experiences 4 consecutive outliers, it’s morelikely that the next sample will be normal than it will stillbe an outlier. We can remark that this observation holdsfor all cases: we have a strictly larger probability to have kconsecutive outliers than k+1. In other words, outliers havea limited impact and the average and mean values would bewell-estimated if they are properly detected and discarded(i.e. they will not introduce a large bias).

Number of outliers per sliding window (WS=20)

Pro

ba

bili

ty

!

!

!!

! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

0 1 2 3 4 5 6 7 8 9 11 13 15 17 19

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

Figure 9: CDF of number of outliers per sliding win-dow of size 20 samples

5.4 Number of outliers in a sliding windowA problem occurs when we want to detect outliers prac-

tically. It is almost impossible to keep the whole history ina node memory in order to precisely calculate IQR and toaccurately remove outliers. We assume we may only savethe last few samples. We chose here to implement a slid-ing window of 20 samples. Further, we have calculated theCumulative Distribution Function of the number of outliersper sliding window (Figure 9) to justify our choice.

In 97.5% of cases, we have 4 or less outliers per slidingwindow (cf. the dashed line). We aim here at limiting thememory consumption while still well estimating the aver-age value to be able to accurately detect the outlier values.Moreover, the method must not be too conservative sincetestbeds are not static and the environment can change. Inparticular, the quality should sometimes be re-estimated,even if new values are far from the previous average values.

We propose the following approach to reach this objective.At most 4 slots will be used to store outlier values (yellowfields in Figure 10). These values will be flagged and won’tbe used to calculate the IQR value (eq. 1) since we considerthat these values are abnormal. Possibly, a new value couldbe detected as outlier although 4 values were already flaggedoutliers. In this case, we un-flag the outlier closest to the

X X X

Q1 Q3

IQR

4

20

4

X

Figure 10: Sliding window example with 4 valuesflagged as outliers

median value. IQR are updated and possibly the outlierscould be considered as normal if they lie, after the update,in the correct range.

Let’s consider Figure 10. We can see that the extreme 4values on the right are flagged as outliers and thus are notused to compute IQR, Q1, etc.

Using this approach, we will smooth the quality metric anddiscard inaccurate measures. Moreover, we are also reactive:we e"ciently detect changing radio links and update theirassociated quality metric accordingly.

In conclusion, if each value can be coded in sizesample, anode only has to reserve 20 $ sizesample bits to compute anaccurate average metric.

The reader can note that such a statistical approach couldbe easily applied to any metric measuring the quality of aradio link.

6. RELATED WORKAfter working on theoretic protocol issues and energy sav-

ings in WSN, researchers in the last years have taken upexperimentally studying the behavior of WSNs [18, 19]. In-deed, simulations are often too simplistic and new problemscan arise in experiments that would have not been appearedin simulations. Barren et al. provide guidelines to conceiveWSN testbeds [6].

Since experiments require much e!ort, some platforms aimat simplifying the performance evaluation of protocols andalgorithms. Orbitlab [5] deployed a grid of sensors in an iso-lated environment. Some other approaches have been stud-ied for more finely controlling a testbed and obtaining repro-ducible experiments [20, 21, 22]. Lei et al. proposed to mapreal environments into an artificial testbed [23]. All theseapproaches aim at simplifying the deployment by proposinga controlled environment.

Other approaches aim at deploying an ad hoc testbed forevaluating the performances of one particular protocol, e.g.[5] for AODV and OLSR, [24] for network coding, [25] for arouting metric, etc. Their purpose is specific and researchershave a hard time trying to generalize these results.

Recently, Raman et al. studied the problem of interferencein ieee 802.11 wireless mesh networks and tried to find outif the concept of a radio link is valid in this kind of networks[26]. Their results are experimental, but lead to fundamentaland generic results in wireless mesh networks. Although itis focused on mesh networks, it gives an overview of whatconcerns may raise in Wireless Sensor Networks.

Some authors conducted real-world experiments. In par-ticular, Barren et al. deployed a WSN to obtain meteoro-logical data [27]. These measures will have been later usedto derive meteorological models and to predict floods.

Ceriotti et al. deployed a WSN for measuring the move-ments of the Torre Aquila [28]. The authors provided feed-back on the tools an engineer should deploy to design ane"cient WSN. Both articles demonstrated that precise mea-sures sometimes cannot be obtained from a single point ofmeasurement. Distributed systems are more robust and per-mit to obtain rich information. However, they do not pro-vide insights into the WSN environment itself. Thus, weproposed here to fill in this gap.

7. CONCLUSION & PERSPECTIVESIn this paper, we have proposed a way to extend the use

and the contribution of implemented testbeds in a urbanenvironment. We have carried out thorough statistical anal-ysis on a collected dataset to obtain an insight on the WSNenvironment and to emphasize its most distinguished prop-erties.

Our analysis considered the aspects of WSNs such as thelink characterization, correlation with environmental param-eters (temperature, humidity, and luminosity) as well as net-work dynamics.

First, we showed that, contrary to the literature, therewere no unidirectional links in our observed testbed, andmore over that all bidirectional links are highly symmetri-cal while comparing their mean RSSI values. Furthermore,we have shown that RSSI values don’t follow any of thebasic distribution models (Normal-Gaussian, Logistic andCauchy): an accurate distribution still has to be proposed.

Although it is well-known that high humidity may causedecrease in link quality, we have shown that there is no corre-lation between humidity and RSSI in our experiments. Evenextreme maximum values of humidity don’t cause significantchanges in link quality measurements. This is most proba-bly thanks to the MAC and PHY layers present in the radiochips.

We have highlighted that a proactive approach in neigh-borhood discovery may cause imprecise routing decisions,favoring the reactive solutions. Besides, although the RSSIexhibits large variations and does not correlate well to thelink quality, we were able to characterize the stable links. Inparticular, high RSSI (more than -75 dBm) or a combina-tion of both distance less than 70 m and RSSI between -75and -90 dBm permit to conclude that we face stable links.

Finally, we presented dynamic, reactive but still flexiblemechanism for detecting and discarding transitory outliervalues in measured RSSI value.

Besides, we could now give some advices to researcherswho aim at deploying a testbed for characterizing the WSNenvironments:

• the network should be globally synchronized: we wouldhave been able to compute also average delays, andgive upper bounds on delays;

• the Packet Error Rate should be included for bothneighborhood discovery packets and data packets inorder to obtain a second metric of reliability;

8. ACKNOWLEDGMENTSThe authors would like to thank the personnel of Orange

Labs (especially Giyyarpuram Madhusudan) for providingthe experimental data.

This work was supported in part by the French Gov-ernment and the Competitive Clusters Minalogic and Sys-tem@tic under contracts FUI SensCity and ANR-09-VERS-017.

9. REFERENCES[1] J.N. Al-Karaki and A.E. Kamal. Routing techniques

in wireless sensor networks: a survey. IEEE WirelessCommunication, 11(6):6 – 28, December 2004.

[2] Giuseppe Anastasi, Marco Conti, Mario Di Francesco,and Andrea Passarella. Energy conservation inwireless sensor networks: A survey. Ad Hoc Networks,7(3):537–568, 2009.

[3] David Cavin, Yoav Sasson, and Andre Schiper. On theaccuracy of manet simulators. In POMC, pages 38–43,Toulouse, France, October 2002. ACM.

[4] Mineo Takai, Jay Martin, and Rajive Bagrodia.E!ects of wireless physical layer modeling in mobilead hoc networks. In MobiHoc, pages 87–94. ACM,Long Beach, CA, USA, 2001.

[5] D. Rastogi, S. Ganu, Y. Zhang, W. Trappe, andC. Gra!. A comparative study of aodv and olsr on theorbit tesbed. In Milcom, Orlando, Florida, USA,October 2007. IEEE.

[6] Guillermo Barrenetxea, Francois Ingelrest, GunnarSchaefer, and Martin Vetterli. The hitchhiker’s guideto successful wireless sensor network deployments. InSensys, pages 43–56, Raleigh, USA, November 2008.ACM.

[7] Anukool Lakhina, Konstantina Papagiannaki, MarkCrovella, Christophe Diot, Eric D. Kolaczyk, and NinaTaft. Structural analysis of network tra"c flows. InSIGMETRICS, pages 61–72, New York, NY, USA,2004. ACM.

[8] Haakon Ringberg, Augustin Soule, Jennifer Rexford,and Christophe Diot. Sensitivity of pca for tra"canomaly detection. SIGMETRICS PerformanceEvaluatation Review, 35(1):109–120, 2007.

[9] Thomas Watteyne, David Simplot-Ryl, IsabelleAuge-Blum, and Mischa Dohler. On using virtualcoordinates for routing in the context of wirelesssensor networks. PIMRC, 2007.

[10] http://www.wavenis-osa.org.[11] Thomas Watteyne, Isabelle Auge-Blum, Mischa

Dohler, Stephane Ubeda, and Dominique Barthel.Centroid virtual coordinates - a novel near-shortestpath routing paradigm. Computer Networks.,53(10):1697–1711, 2009.

[12] Tereus Scott, Kui Wu, and Daniel Ho!man. Radiopropagation patterns in wireless sensor networks: newexperimental results. In IWCMC, pages 857–862,Vancouver, British Columbia, Canada, 2006. ACM.

[13] R.P. Liu, Z. Rosberg, I.B. Collings, C. Wilson, A.Y.Dong, and S. Jha. Overcoming radio link asymmetryin wireless sensor networks. In PIMRC, pages 1–5.IEEE, 2008.

[14] Lifeng Sang, Anish Arora, and Hongwei Zhang. Onlink asymmetry and one-way estimation in wirelesssensor networks. ACM Transactions on SensorNetworks, 6(2):1–25, February 2010.

[15] Kannan Srinivasan and Philip Levis. RSSI is under

appreciated. In EmNets, Cambridge, MA, USA, May2006.

[16] Peter. J. Huber. Robust Statistics. Wiley, 2004.[17] K. Kaemarungsi. Distribution of wlan received signal

strength indication for indoor location determination.In ISWPC, Phuket, Thailand, January 2006. IEEE.

[18] Adel Aziz, Alaeddine El Fawal, Jean-Yves Le Boudec,and Patrick Thiran. Aziala-net: Deploying a scalablemulti-hop wireless testbed platform for researchpurposes. In MobiHoc S3, New Orleans, USA, May2009. ACM.

[19] R. Riggio, D. Miorandi, I. Chlamtac, N. Scalabrino,E. Gregori, F. Granelli, and Yuguang Fang. Hardwareand software solutions for wireless mesh networktestbeds. IEEE Communications Magazine,46(6):156–162, June 2008.

[20] Juan Flynn, Hitesh Tewari, and Donal O’Mahony. Areal-time emulation system for ad hoc networks. InCNDS, pages 115–120, San Antonio, USA, January2002.

[21] James T. Kaba and Douglas R. Raichle. Testbed on adesktop: strategies and techniques to supportmulti-hop manet routing protocol development. InMOBIHOC, pages 164–172, Long Beach, USA,October 2001. ACM.

[22] Sagar Sanghani, Timothy X. Brown, ShwetaBhandare, and Sheetalkumar Doshi. EWANT: Theemulated wireless ad hoc network testbed. In WCNC,New Orleans, USA, March 2003. IEEE.

[23] J. Lei, R. Yates, L. Greenstein, and Hang Liu.Mapping link snrs of real-world wireless networks ontoan indoor testbed. IEEE Transactions on WirelessCommunications, 8(1):157–165, January 2009.

[24] Sachin Katti, Hariharan Rahul, Wenjun Hu, DinaKatabi, Muriel Medard, and Jon Crowcroft. Xors inthe air: Practical wireless network coding.IEEE/ACM Transactions on Networking,16(3):497–510, June 2008.

[25] John Bicket, Daniel Aguayo, Sanjit Biswas, andRobert Morris. Architecture and evaluation of anunplanned 802.11b mesh network. In MOBICOM,Cologne, Germany, August 2005. ACM.

[26] Bhaskaran Raman, Kameswari Chebrolu, DattatrayaGokhale, and Sayandeep Sen. On the feasibility of thelink abstraction in wireless mesh networks. IEEETransactions on Networking, 17(2):528–541, April2009.

[27] G. Barrenetxea, F. Ingelrest, G. Schaefer, M. Vetterli,O. Couach, and M. Parlange. Sensorscope:Out-of-the-box environmental monitoring. In IPSN,St. Louis, Missouri, USA, April 2008. IEEE/ACM.

[28] Matteo Ceriotti, Luca Mottola, Gian Pietro Picco,Amy L. Murphy, Stefan Guna, Michele Corra, MatteoPozzi, Daniele Zonta, and Paolo Zanon. Monitoringheritage buildings with wireless sensor networks: Thetorre aquila deployment. In IPSN, San Francisco,USA, April 2009. ACM/IEEE.